How the World Bank Is Building AI Infrastructure for Global Development

“There is a tendency to think that if you’re not capable of building a frontier model, then you don’t have a voice in the ecosystem.”

Samuel Fraiberger, Founder and Lead of the AI Lab, Development Impact Evaluation Group at the World Bank

Artificial intelligence is quickly becoming a central driver of economic development. Governments around the world are exploring how AI can improve public services, inform policy decisions, and address complex challenges such as food security and healthcare delivery. Yet the ability to use artificial intelligence effectively depends on something far less visible than the models themselves: the infrastructure that allows these systems to function.

Data systems, connectivity, and technical expertise determine whether governments can actually deploy AI tools in practice. In many low- and middle-income countries, these foundations remain incomplete. Reliable statistical data may be limited, internet infrastructure uneven, and local technical capacity constrained. As a result, the question shaping the next phase of AI adoption is not simply who builds the most advanced models, but which countries can build the institutional and technological systems required to use them.

Much of the global conversation around artificial intelligence focuses on frontier models, corporate competition, and regulatory debates in advanced economies. These discussions are important, but they only capture part of the story. For many governments, the more immediate challenge is translating AI capabilities into practical tools to support policymaking and public service delivery.

Development institutions are increasingly experimenting with how to address this gap. Artificial intelligence can function as operational infrastructure for governments, helping policymakers analyze complex data, anticipate risks, and improve the delivery of social programs. Building these systems requires combining technical expertise with development knowledge, a task that few institutions have historically attempted.

Inside the World Bank, one such effort is underway within the Development Impact Group. An applied AI lab within the group is exploring how AI tools can serve as global public goods for governments, particularly in countries where reliable data and analytical capacity remain limited. 

To better understand how artificial intelligence can be developed and deployed as a public good in low- and middle-income countries, we spoke with Samuel Fraiberger, Data Scientist and AI Program Lead in the World Bank’s Development Impact Group.

Sam Fraiberger is the founder and lead of the AI Lab in the Development Impact Evaluation (DECDI) group at the World Bank and an affiliate researcher in the Computer Science Department at New York University. His work sits at the intersection of AI and global development, with research spanning food security, online violence, and AI-augmented evidence synthesis. He leads the development of AI public-good products derived from this research and used by governments, international organizations, and civil society worldwide. 

His research has appeared in leading journals and conferences including Science, PNAS, Science Advances, ACL, and EMNLP. He also has been featured in The Economist, The Washington Post, and The Wall Street Journal. He was selected for the inaugural Google.org Generative AI Accelerator and named one of the Top 100 People in AI in Government by Apolitical. He holds a PhD from New York University and an SM in applied mathematics from Harvard University.

Fraiberger’s work sits at the intersection of artificial intelligence research and development policy. Trained as an AI scientist, he joined the World Bank nearly a decade ago and began building a technical research initiative within the institution’s Development Impact Group. The goal was to explore how artificial intelligence could support governments facing complex development challenges.

“I’m an AI scientist by training,” Fraiberger explains. “I joined the World Bank about eight years ago, and I had the opportunity to create this lab within the research department, which is an applied AI lab where we essentially build or train AI models that are used to build global public goods.”

The lab operates differently from many private sector AI initiatives. Instead of focusing on commercial products or frontier models, the team develops tools intended to support governments and international organizations working in low- and middle-income countries. These systems are designed to help policymakers interpret large amounts of information and respond to challenges across sectors.

“These systems serve governments and partners in low- and middle-income countries to address global challenges,” Fraiberger says. The projects span a range of issues, from food security and online harms to broader efforts to improve decision-making within public institutions. “That really focuses on addressing challenges across sectors from food security to hate speech and online violence, but also more generally to help policymakers, decision makers, and practitioners in low- and middle-income countries make better decisions using AI tools.”

The lab’s creation reflects Fraiberger’s own motivation to combine technical research with real-world impact. While artificial intelligence research often advances in academic or commercial environments, he was drawn to the opportunity to apply those tools to development problems.

“I wanted to do something that was technical but also had a purpose and a social impact,” he says. “One thing leading to another, I ended up building this lab thanks to external funders that helped me grow a team and lead multiple projects across different sectors.”

Today the team brings together engineers, economists, and researchers working on applied AI systems that address practical development challenges. Many of these projects focus on a common problem facing policymakers across the developing world: how to measure rapidly changing conditions when reliable data is scarce.

For many governments, one of the most persistent obstacles to effective policymaking is the difficulty of measuring what is happening in real time. Traditional data sources such as national surveys or census programs are often expensive, slow to produce, and conducted infrequently. In countries where administrative systems are still developing, policymakers may be forced to make decisions using information that is months or even years out of date.

Artificial intelligence is beginning to change that dynamic by enabling researchers to analyze data sources that were previously too complex or fragmented for policymaking. Fraiberger describes measurement as one of the most important areas where AI can expand the capabilities of development institutions.

“Measurement is a large area where the World Bank and other organizations have been playing an important role,” he says. Much of that work involves incorporating what researchers call non-traditional data sources into development analysis.

“That really has to do with any type of unstructured data like remote sensing, satellite imagery, or data from social media and others that are used to extract insight about what’s going on in countries where it’s hard to necessarily have good measures,” Fraiberger explains.

These sources can provide signals about social and economic conditions that are difficult to capture through traditional statistical methods. Satellite imagery can reveal patterns in agricultural production or urban growth. Social media and online content can offer clues about public sentiment or emerging crises. When analyzed with machine learning models, these signals can help governments understand conditions on the ground much faster than conventional data collection methods allow.

Recent advances in generative AI have accelerated the development of these tools. According to Fraiberger, tasks that once required extensive manual data processing can now be completed far more efficiently.

“That work has been tremendously accelerated with generative AI because a lot of the models and products that we build are so much more efficient now to develop,” he says. “Lots of the processes are built much faster than they used to.”

One of the lab’s most prominent initiatives illustrates how these ideas translate into practical tools. The project, known as Zero Hunger AI, focuses on predicting food insecurity across dozens of countries by analyzing large volumes of publicly available information.

The concept grew out of Fraiberger’s academic research examining how news reporting reflects conditions on the ground during emerging crises. In a study published in the journalScience Advances, he and his collaborators found that global news coverage often contains early signals of food security risks.

“If you look at the information published by journalists around the world about what’s going on, it tells you a lot about risks on the ground like pest infestation or climate shock,” he says.

Despite the richness of this information, it has historically remained outside the formal tools used by policymakers.

“All this data is very rich and very informative, but it’s not really used for public policy and decision-making,” Fraiberger explains.

By applying artificial intelligence to these information streams, the team developed models that can detect patterns across large volumes of reporting and translate them into early warning indicators. These signals can then be incorporated into systems that help governments and international organizations anticipate food crises before they escalate.

“What we figured out is that using AI we could extract valuable signals from these data sources and embed them into products to inform governments, NGOs, and other multilaterals,” Fraiberger says.

The resulting forecasts can support programs designed to respond to food insecurity earlier and more effectively. In some cases, this information helps guide anticipatory cash transfer programs that provide financial assistance to vulnerable populations before crises deepen.

For Fraiberger, the broader significance of projects like Zero Hunger AI lies in how they reshape the role of data in development policy. Instead of relying solely on slow statistical reporting cycles, governments can begin to integrate a wider array of information sources into their decision-making processes.

This shift has implications far beyond food security. As AI tools improve, similar approaches may help policymakers monitor economic shocks, identify emerging public health risks, or track social instability in regions where traditional data remains limited.

Yet the ability to build and deploy these systems also raises deeper questions about the infrastructure required to support them.

Projects like Zero Hunger AI demonstrate how artificial intelligence can expand the tools available to policymakers. At the same time, deploying these systems across diverse development contexts reveals the structural constraints that shape how AI can be used in practice.

Fraiberger identifies three core inputs that determine whether artificial intelligence systems can function effectively in low- and middle-income countries.

“The three big inputs in developing AI products in low-income contexts are connectivity, data, and expertise,” he says.

Connectivity remains one of the most immediate barriers. Many regions still lack reliable internet infrastructure, which complicates the deployment of cloud-based AI systems. In these environments, teams often need to design tools that can function with limited connectivity or operate offline.

Data availability presents another challenge. Many artificial intelligence models rely on large volumes of structured training data, which may not exist in many development contexts. Addressing this gap requires researchers to adapt models, create new datasets, and develop methods for working with incomplete information.

“We do a lot of work on fine-tuning and on creating our own data for ensuring that the products are very high quality,” Fraiberger explains. The process often involves building datasets from scratch or adapting existing models, so they perform reliably in new environments.

The third constraint involves technical expertise. Artificial intelligence systems require specialists capable of developing, maintaining, and evaluating complex models. Building this capacity within governments and local institutions is essential if these tools are to be used sustainably.

“Knowledge transfer is really something that can help countries leapfrog,” Fraiberger says, emphasizing the importance of training and collaboration in expanding local technical capacity.

Developing these systems inside a large international organization introduces additional complexity. Institutions like the World Bank have historically focused on research outputs such as reports, policy analysis, and statistical datasets. Producing operational AI systems represents a relatively new type of activity.

“There is no blueprint on how institutions like the World Bank need to provide AI public goods for low- and middle-income countries,” Fraiberger says.

The technology itself is evolving rapidly, which means institutional processes must adapt as well. Building AI tools within large organizations often requires navigating internal procedures designed for more traditional forms of research and development.

“Moving the frontier in an environment that can be a little bureaucratic can sometimes be challenging,” he explains.

Despite these challenges, the effort reflects a broader transformation in development practice. Artificial intelligence is increasingly becoming not only an object of policy analysis but also a practical tool that institutions must learn to build and deploy.

As these systems spread, however, the conversation about artificial intelligence is also beginning to raise larger questions about governance, equity, and who ultimately benefits from the global AI economy.

As artificial intelligence becomes a central focus of global technology policy, much of the public debate has concentrated on frontier risks. Discussions often center on the regulation of advanced models, the possibility of catastrophic misuse, and competition among major technology companies.

Fraiberger believes these concerns are important but incomplete. Many of the most immediate challenges surrounding AI are emerging in contexts that receive far less attention.

“We hear a lot about regulation of AI in the West and catastrophic risk or different kinds of harms that frontier models could lead to,” he says. “But I think it’s equally important to think about the harms that these systems can lead to in contexts in which maybe there is less spotlight and their impact is less visible.”

One example involves language and cultural representation in AI systems. Many large language models perform well in English and other widely represented languages because of the large volumes of training data available. In languages with fewer digital resources, performance can degrade significantly.

“The systems obviously perform quite well in the tasks that we do when it’s in English and when there’s a lot of data to train on,” Fraiberger explains. “When you try to do the same thing in rare languages or in contexts that these models don’t have as much knowledge about, they tend to perform much worse.”

These limitations are not only technical. In environments where resources for testing and oversight are limited, poorly performing systems can spread misinformation, reinforce bias, or create instability.

For Fraiberger, addressing these issues requires a more systematic approach to evaluating how AI systems perform across different contexts.

“One of the things I’m very strongly advocating for is systematization of benchmarks,” he says. Just as frontier models are regularly tested on coding or reasoning benchmarks, similar evaluation frameworks could be developed to measure performance across languages, cultures, and social environments.

Another issue that receives little attention in global AI debates involves the labor that supports the development of large AI systems. Training large language models requires vast quantities of labeled data, much of which is produced through annotation work often carried out in developing economies.

“There is not a lot of transparency on where people providing labor in annotation work for large language models are located, or how much they’re compensated,” Fraiberger says.

These workers contribute directly to the creation of technologies valued at billions of dollars.

“Workers are providing labor to build products that are then valued for billions of dollars without being fairly compensated,” he explains.

In Fraiberger’s view, countries should think more strategically about their role in the global AI ecosystem. Even if they are not building frontier models themselves, they still possess assets that are critical to the development of artificial intelligence.

“There is a tendency to think that if you’re not capable of building a frontier model then you don’t have a voice in the ecosystem,” he says. “But there is actually a lot of assets that come from data work that’s being put into building frontier models.”

Those assets include training data, digital information produced by national economies, and the labor that supports data annotation and model development. Governments could use these resources more strategically, for example through regional cooperation or new regulatory frameworks.

“Countries have to think really strategically about their assets,” Fraiberger says, suggesting that data and labor contributions could be treated as economic resources rather than passive inputs into global technology systems.

The future of artificial intelligence will likely be shaped not only by the companies that build the most advanced models, but also by the institutions and countries that determine how those models are used. For many governments, the challenge is not simply adopting AI tools, but ensuring that they participate meaningfully in the systems that define the next generation of technological development.